Y Combinator's Garry Tan Calls for American AI Model Distillation Regime Amid China Accusations

Y Combinator CEO Challenges Distillation Restrictions

Garry Tan, the Y Combinator CEO, has taken a contrarian stance on AI model distillation, arguing that U.S. regulators should permit American open-weight AI labs to use the same techniques Chinese competitors employ. Speaking at Y Combinator's annual Demo Day, Tan told CNBC he would "do nothing" about distillation concerns, instead proposing "an American distillation regime" that gives the U.S. more robust American open-weight AI alternatives2

2

.

This position directly contradicts recent calls from Anthropic CEO Dario Amodei, who urged regulators to crack down on what the company calls "illicit distillation attacks." Anthropic released its second report this week alleging Chinese labs hide their identities to distill without permission, relying on fraud and stolen credentials1

1

. The U.S. National Security Agency, Cybersecurity and Infrastructure Security Agency, and Federal Bureau of Investigation issued an official cybersecurity advisory warning on Tuesday2

2

.

Distillation Debate Intensifies Among AI Giants

AI model distillation involves extensively prompting one model to learn how another works and reasons. While commonly used legitimately by AI labs to train new models, AI giants accuse China of crossing ethical lines. Anthropic has specifically named Chinese companies including Moonshot AI, DeepSeek, and MiniMax, while OpenAI believes DeepSeek's V3 and R1 architectures were distilled from its GPT-4 and GPT-4o models2

2

.

Tan clarified he doesn't advocate for American labs to use stolen credentials. Instead, he wants them free to come through the front door. His argument is twofold: proprietary AI labs shouldn't dictate what customers do with information their models share, and these same labs didn't ask permission when they vacuumed up copyrighted material for training data1

1

.

Balancing Open-Weight AI and Frontier AI Systems

"Controlling what users and customers do with API calls to closed weight models feels constraining," Tan told TechCrunch. He believes access to intelligence trained on broad public data should be "more a form of a public good than something locked away behind restrictive terms of service"1

1

.

Tan envisions a balance between open-weight AI and frontier AI systems, acknowledging that frontier models drive innovation forward and need sustainable business models. He told CNBC this requires maintaining a price premium that keeps proprietary AI labs viable while ensuring open-weight models provide freedom and access. "This is actually the ideal case," he explained, calling it "a tightrope" but "a balance worth pursuing"2

2

.

Warning Against Monopolistic Scenario

Tan's greatest concern isn't distillation but concentration of power. "The nightmare scenario, the doomer scenario for AI is that there's just one company," he said. "It has the best access to capital. It has the best AI researchers. It runs away with it and suddenly there's one company that's monolithic. And that would be bad"1

1

.

Focusing on Near-Term AI Misuse Over Doomsday Scenarios

Tan advocates focusing on immediate risks rather than doomsday-style AI risks currently dominating headlines. "We need to be focused on science fact, not science fiction," he stated. "We need to be responding to what is happening right now"2

2

.

He identified cybersecurity as an imminent threat and pointed to near-term AI misuse cases. On Thursday, Anthropic reported blocking Claude access for five cases where scientists in unspecified foreign countries were using the models for bioweapon research on dangerous pathogens2

2

.

Regarding job displacement, Tan sees a decades-long timeline rather than immediate impact. He foresees people transitioning toward automating rote tasks while spending more working hours on creative pursuits. "It will take decades for this to actually percolate into society, and that's not a bad thing," he said2

2

.

At Y Combinator's Demo Day, 149 of 196 startups presenting were categorized as machine-learning and AI ventures, demonstrating the accelerator's continued investment in the space2

2

.